Semi-Supervised Approach for Automatic Counting of Whiteflies With Small Annotated Dataset

Insect counting is key action for pests’ control in agriculture. Automatic insect counting would allow a fast and accurate characterization of the infestation degree which would lead to a better choice of insecticide dose and, consequently, more effective treatments. Recently, an approach that automatically counts the insects in the wild has been proposed by Bereciartua-Pérez et al. That method is based on density map estimation with deep learning and has proven to offer very good results. Deep learning techniques, however, still present one big drawback: they rely on lots of annotated data. In the case of insect counting by density map estimation, the annotation process is a very tedious and time-consuming task and it entails an important bottleneck in the development of the model. In this paper, a new semi-supervised method is proposed for automatic counting of whiteflies with a small annotated dataset. Semi-supervised learning is based on leveraging not annotated data during training. Our semi-supervised method is based on the design and implementation of a pseudo-annotation algorithm that requires few annotated data. The pseudo-annotations obtained from this algorithm might be noisy but they help during the training of the whitefly counting model allowing to reduce the manual annotations needed and, therefore, reducing the effort and time needed to get a usable deep learning based solution for the task. Our new semi-supervised approach using only 48 manually annotated images achieves similar results as the fully supervised approach trained with 474 manually annotated images.